Knowledge Engineering for Automated Planning

1
Knowledge Engineering for Automated Planning

• Lee McCluskey
• With acknowledgement to
• Ron Simpson

2
Abstract

• Algorithms for automated planning have become
  significantly more powerful in recent years, but
  the industrial take up of the technology has been
  slow. One reason for this has been the difficulty
  of modelling problem areas with sufficient rigour
  to allow for their automated solution.
• The tutorial will provide attendees with
• an insight into the range of potential
  applications of the technology
• an idea of the level of difficulty in deploying
  the technology
• Information on where to find out more and how to
  obtain free software

3
Abstract

• The tutorial will cover
• a brief introduction to automated planning from
  an AI perspective (AI Planning)
• a review of the scope and power of the currently
  available automated planning systems.
• the languages used for domain problem
  specification and the associated knowledge
  engineering issues.
• the development of example domain models using
  dedicated knowledge engineering tools in GIPO.
• practical demonstrations!

4
Introduction
5
Introduction Resources

• Mainly because of PLANET (EC F5 NoE in Planning)
  there are heaps of resources to back up this
  tutorial
• Applications of AI Planning website
• http//vitalstatistix.nicve.salford.ac.uk/planet2
  /
• Free Planners, Schedulers and Domain Models
• http//scom.hud.ac.uk/planet/repository/
• General PLANET website containing Summer School
  Notes on many aspects of AI Planning, and
  Planning Curriculum
• http//www.planet-noe.org/

6
Introduction Resources

• Free downloadable software GIPO - to engineer
  planning domain models
• http//scom.hud.ac.uk/planform/gipo/
• KE for Planning ROADMAP
• http//scom.hud.ac.uk/planet/home/
• The notes will be on my website
• http//scom.hud.ac.uk/scomtlm

7
Introduction What is AI Planning?

• The scope of AI Planning is the synthesis
  (generation) and execution of PLANS. That is, AI
  Planners reason with actions and generate plans.
• Scheduling is the allocation of resources to
  plans of action taking into account various
  constraints generally considered an easier
  subpart of the Planning process
• Planning is considered more of a knowledge-based
  pursuit than scheduling

8
Introduction MAIN ASSUMPTION

• The MAIN ASSUMPTION of virtually all work in AI
  Planning is that there should be a logical
  separation between
• The Planning Engine
• The Domain Model

Planning System
Planning Engine
APPLICATION DOMAIN
Domain Model
9
Introduction

• Implications of assumption
• It tends to made AI Planning a different subject
  to eg planning and scheduling in manufacturing
• planning engines AND domain models can be
  developed, tested, debugged, validated
  independently - the model of the particular
  application of planning is developed in relative
  isolation to the planning engine
• domain models may be useful for more purposes
  that simply automated planning functions

10
Introduction

• Negative consequences of assumption ..
• Domain model representation is influenced by
  what planners an handle
• Consequent inefficiency in planning systems -
  akin to compilation rather than hand coding
  in software
• Much research work has been carried out in
  developing planning engines, pushed along by the
  International Planning Competition (AIPS98,
  AIPS00, AIPS02, ICAPS04) BUT methods and tools
  to help develop the domain models have had
  relatively little attention

11
Introduction Applications of AI Planning

• See web site http//vitalstatistix.nicve.salford.a
  c.uk/planet2/
• Aerospace e.g.
• Autonomous control of spacecraft - the NASA
  Remote Agent experiment
• Vicar planner applied to automated image
  processing
• Satellite mission planning and scheduling
  (European Meteostat)
• Planning and Scheduling of Spacecraft Assembly

12
Introduction Applications of AI Planning

• Military e.g.
• Planning for military air campaigns
• DARPA/Rome Planning Initiative - A military
  scheduling project

13
Introduction Applications of AI Planning

• See web site
• http//vitalstatistix.nicve.salford.ac.uk/planet2/
  
• Industrial / Government / Control e.g.
• Planning in a forest fire simulation system -
• .. disaster recovery in general
• Brewery production-line scheduling
• Generation of control programmes for industrial
  plant
• Oil Spill Response Planning
• Ship Building using shipyard scheduling
  optimisation systems
• Aircraft crew scheduling
• Chemical Plant control

14
Introduction Applications of AI Planning

• Other Current and Future Application areas
• Other areas in Control
• Workflow/Workforce
• Air Traffic
• Project planning
• Transport Logistics
• Web Agents,
• Games Software
• Physical Robots, Autonomous Vehicles

15
Basic Concepts in AI Planning
16
Basic Concepts in AI Planning

• Basic concepts in AI Planning are
• Operators, Objects, States, Goals, Planning
  Problem, Plans
• To illustrate them we will use a famous benchmark
  domain the change tyre world

17
Basic Concepts in AI Planning

• A state represents a world or snapshot within the
  domain. Objects have States the conjunction of
  all objects states make up a world state
• wrench_in(wrench1,boot)
• wheel_in(wheel1,boot),pumped_up(wheel1)
• wheel_on(wheel2,hub1),flat(wheel2)
• pump_in(pump1,boot)
• tight(nuts1,hub0)
• have_jack(jack1)
• on_ground(hub1),fastened(hub1)
• closed(boot)

18
Basic Concepts in AI Planning

• An operator represents an action within the
  application. Operators change state.
• operator(putaway_wheel(C,W),
• prevail
• se(container,C,open(C)),
• necessary
• sc(wheel,W,have_wheel(W)gtwheel_in(W,
  C)),
• conditional
• )
• A Domain model consists of a set of operators

19
Basic Concepts in AI Planning

• Goals are conditions on states
• wheel_on(wheel1,hub1)

20
Basic Concepts in AI Planning

• A Planning Problem is a triplet
• (Initial World State, Goal, Domain Model)
• Plans are collections of (instantiated, ordered)
  operators that solve planning problems

21
Basic Concepts in AI Planning

• Other concepts in AI Planning are
• Tasks, Events, Resources, Time
• - Tasks are compound actions to be executed (eg
  make a cup of tea)
• Events change the states of objects but happen
  exactly when their pre-conditions are true
• Resources are quantities used up by actions -
  usually represented by numeric variables

22
Basic Concepts in AI Planning

• How do AI planners generate plans?
• well there are MANY techniques ..
• Genarally they SEARCH through some representation
  of the planning problem!

23
Current State of AI Planning
24
Current State of AI Planning

• Research in AI planning has produced very good
  results in the last 10 15 years largely due
  to
• The International Planning Competition
• Acceptance of a standard communication language
  called PDDL for domain models and domain problems
• More industrial involvement NASA US military

25
OLD State of AI Planning

• In the 80s / early 90s.. Most planners could
  solve simple problems with domain models
  consisting of
• Actions modelled as instantaneous, deterministic
  operators with infinite resources.
• Actions pre-conditions and effects were
  propositions
• States set of propositions under the CWA
• Goal set of propositions.
• Metrics for planners
• time to solve problem ie generate plan
• size (no of operators) in sequential plan

26
Current State of AI Planning

• Now Planners can solve more complex problems
  with domain models consisting of ..
• Durative operators time is explicit
• Resources
• Non-deterministic operators
• Operators with complex/conditional effects
• Partially observed states
• More metrics considered e.g. makespan and
  multi-objective achievement
• MOST IMPORTANT they are downloadable!

27
Current State of AI Planning

• PDDL is the common communication language. Main
  variants of PDDL..

PDDL 1 1998 first IPC
2002 - Added Duration and Numerical Quatities
PDDL 2.1
2003 - Added timed initial facts, derived
predicates
PDDL 2.2
PDDL
2002 - Added Processes, Events, cts time
28
Current State of AI Planning

• Now many very effective techniques in use in
  plan generation..
• (see http//scom.hud.ac.uk/planet/repository/ )
• Generate and search through a plan graph
  (Graphplan, STAN)
• Do best-first, forward, state space search with
  very good weak heuristics (FF, HSP)
• Compile planning problem into a large set of
  clauses and solve with a satisfiability engine
  (Blackbox)
• Compile planning problem into a compact storage
  form such as Binary Decision Diagrams (MIPS)

29
Current State of AI Planning

• Summary - Good News
• Plan generation algorithms are much more
  efficient than 10 years ago, and can work
  efficiently in more expressive problems domains
• But there is Bad News
• Technology transfer there is much to do in
  making the technology generally available and
  usable
• Model development domain model authors use
  planners themselves to try to develop and debug a
  domain model. Planners have not generally been
  designed for this purpose..

30
Knowledge Engineering and Domain Model Capture
31
Knowledge Engineering

• Knowledge Acquisition / Engineering is a huge
  area in AI related to Knowledge-based Systems
  (KBS)
• Old idea of 'knowledge transfer', where
  constructing a KBS amounted to extracting the
  knowledge from experts and encoding it within an
  expert system 'shell (20 years ago!)

Application expertise
transfer
Procedural expert knowledge
32
Knowledge Engineering

• Now KBS emphasises the building of a deep causal
  model prior to an operational system.
• This domain model has to embody not just the
  procedural expert knowledge but the environment
  in which this knowledge was utilised.
• Several modelling frameworks have been developed
  (e.g. CommonKads which is based on the use of a
  series of models during domain capture, each
  dealing with different aspects of the domain.)
• These support the process of model acquisition
  and validation, and are underpinned by an overall
  method of development.

33
Knowledge Engineering for AI Planning Definition

• Knowledge Engineering (KE) in AI Planning is the
  process that deals with
• acquisition,validation and maintenance of
  planning domain models, and
• the selection and optimization of appropriate
  planning machinery to work on them.
• Hence, knowledge engineering processes support
  the planning process they comprise all of the
  off-line, knowledge-based aspects of planning
  that are to do with the application being built.
• (definition of Roadmap http//scom.hud.ac.uk/pla
  net/home/)

34
Knowledge Engineering for AI Planning

35
Knowledge Engineering for AI Planning

• KE for KBS Is generally not the same as KE for
  planning..
• The knowledge elicited in planning is largely
  knowledge about actions and how objects are
  effected by actions. This knowledge has to be
  adequate to allow efficient automated reasoning
  and plan construction.
• The ultimate use of the planning domain model is
  to be part of a system involved in the
  synthetic'' task of plan construction (not for
  solving diagnostic or classification problems as
  in typical KBS)

36
Knowledge Engineering for AI Planning
Terminology

• Domain is the application area
• Domain model is a formal model (theory) of the
  application area
• Acquisition is the process of producing a domain
  model of the application area
• Modelling is the area of using the model to
  predict behaviour in the application area

37
Knowledge Engineering for AI Planning
Terminology

Symbolic World
DOMAIN APPLICATION AREA
Acquisition
Domain Model
Domain Model Language
Predict
Modelling
38
Knowledge Engineering for AI Planning Validation

• Validation of a model is the process that
  promotes its quality in terms of internal and
  external criteria by the identification and
  removal of errors in the model.
• Internal criteria includes properties such as
  syntactic correctness and logical consistency in
  general these properties can be proved formally
  and are not too problematic.
• External criteria includes properties such as
  accuracy, correctness and completeness. Given
  that the sources of the model will not often be a
  mathematical object, these properties can never
  be proved correct (in the same sense that a
  requirements specification can never be proved
  correct).
• Note the distinction between validation of a
  domain model and validation of a planning system.
  The former supports the latter, and occurs at a
  much earlier stage in system development.

39
Domain Model Languages for AI Planning

• A DML should
• Be associated with a method.
• Be tool supported
• Be expressive and customizable
• Support the operational aspects of the model
• Have a clear syntax and semantics
• Be structured

40
PDDL

• PDDL is the standard communication language for
  domain models but
• has no associated method for building models
• has little structure for helping in model
  building
• has little in the way of static tools to help in
  de-bugging

41
OCLh

• OCLh is a language developed precisely for
  helping with the BUILDING of domain models
• It is similar in some respects to PDDL but-
• It has structure using object classes and state
  abstractions
• It has a tools environment called GIPO which
  supports a model building method

42
Planning Domain Engineering with GIPO
43
GIPO - rationale

• Planning Domain Models are hard to design, write,
  debug, maintain - even for experts. The process
  of encoding is laborious. Bugs are of various
  types can lurk in models for a long time.
• As planners and planning applications become
  larger, the problems of engineering planning
  domain models become more acute. There is a need
  to research into engineering environments and
  explore their synergy with general purpose
  planners.

Application
Domain Model
Acquisition is Very hard!!
44
GIPO - rationale

• The two main planning systems used in anger are
• O-Plan (Edinburgh University)
• SIPE (Stanford Research Institute)
• Both have very expressive domain models
  languages.
• To make an application efficient, the user must
  encode appropriate heuristics.
• To be able to use them one has to be a planning
  expert, modelling expert and an application
  expert.
• Even then, developing models is a painstaking
  process.

45
GIPO what is it?

• GIPO (Graphical Interface for Planning with
  Objects) is an experimental GUI and tools
  environment for building planning domain models.
• It is written mainly in Java, with some embedded
  tools in Prolog, and is under continuous
  development.
• It is a product of PLANFORM, a UK EPSRC-funded
  research project, written at The University of
  Huddersfield UK. Website http//scom.hud.ac.uk/p
  lanform
• Our long term aim is make the technology more
  usable and available!

46
GIPO versions

GIPO 1.1
Generally available For Flat models (ECP01)
GIPO
GIPO 1.2
Not on release For models with cts time,
events and processes (PlanSig03)
Not on release Incorporating automated induction
of Operators (AIPS02)
GIPO 2
Generally available For hierarchical
models (ICAPS03)
47
GIPO -functions

• GIPO allows a user to create new domain models or
  import and change old ones via a GUI. It features
• on-line tutorial and OCL manual
• Tools for initial model acquisition
• Tools for model validation
• Planning engines
• 3rd party Planning engines can be easily bolted
  on to GIPO 1 as it outputs PDDL 1.2, and can
  accept generated plans from them.

48
GIPO main tools

• syntax and semantic checks for individual
  components, and between components, of a model
• From Trivial checks on names and sorts
• To complex checks on the structure of
  hierarchically defined operators
• a plan stepper
• a plan animator
• a random task generator (GIPO 1 only)
• an operator induction method (GIPO 1.2 only)

49
GIPO simple method

• Identify objects and object classes (sorts)
• Define predicates
• Define typical object states
• Define operators
• Debug and validate using plan stepper, planner
  and animator
• Method in more detail was given in
• T.L. McCluskey and J.M. Porteous Engineering and
  Compiling Planning Domain Models to Promote
  Validity and Efficiency. Artificial Intelligence
  Vol. 95(1), pages 1 - 65, 1997.

50
Future releases GIPO

• GIPO supports models that contain
• Actions
• Events
• Processes
• Continuously varying values (Time)

51
GIPO Air Traffic Control Example

All Symbols are clickable and give object
information
plane flying through a block represented by a
process
52
Air Traffic Control Example

Safety Violation Event
53
Conclusion

• AI Planning is a maturing technology
• AI Planning technology needs to made more usable
  and available
• The main assumption in AI Planning is that there
  should be a logical separation between the
  planning engine and the domain model.
  Particularly important is the engineering of the
  domain model if this has bugs then the
  application is doomed
• GIPO is an experimental GUI for building and
  validating planning domain models. It is a first
  step in making planning technology more usable.